OpenAI and Anthropic Hunt Smaller Data Centers to Close the Inference Capacity Gap

OpenAI and Anthropic are now shopping for 20-30 MW data center slots alongside their multi-gigawatt campus commitments, a shift that signals just how acute the inference capacity crunch has become. According to a CNBC report, both labs are exploring existing facilities in the UK, Nordics, and US to bring compute online faster while their larger builds remain years from completion.

Both labs are chasing 20-30 MW sites for inference workloads

Anthropic has reportedly sounded out potential deals across the UK and Nordic countries, while OpenAI has explored similar opportunities in the Nordics and the US, according to sources cited by CNBC. OpenAI told CNBC it is building a diversified compute portfolio and evaluates deployments by performance, reliability, timing, and cost. Anthropic declined to comment.

These smaller sites complement the enormous campuses both companies have already committed to. Anthropic reportedly signed a roughly $45 billion agreement with Nscale for around 460 MW of compute at a West Virginia development. OpenAI's Stargate commitments have surpassed 10 GW, with an additional 3 GW in Georgia and 8 GW in Ohio. But those projects take years to build, requiring land, grid connections, substations, and cooling infrastructure before any accelerator goes online.

Why smaller sites matter for inference-heavy workloads

Inference can be distributed across multiple smaller clusters, making 20-30 MW facilities appropriate for serving models to users. Renting capacity from hyperscalers and neoclouds is already standard practice, so this strategy extends an existing pattern. The appeal is speed and location rather than a retreat from long-term builds.

Jabez Tan, head of research at Structure Research, told CNBC that securing a few megawatts at an existing powered site can be more practical than waiting for a much larger block in one location. For builders shipping AI products, distributed inference across multiple sites could reduce latency, improve reliability, and avoid single points of failure. The approach also bypasses some of the local opposition that has blocked 45 data center projects worth $68 billion in Q2 2026 over land use, noise, and water concerns.

What remains uncertain about these deals

The reported talks are still in early exploration, and specifics on deal sizes, locations, and terms could change as negotiations continue. OpenAI's head of data centers, Chris Malone, recently left the company, adding some organizational uncertainty. Meanwhile, the massive campus projects in West Virginia, Georgia, and Ohio remain in development and face their own permitting, grid, and construction timelines.

For now, the key takeaway is that both labs are treating inference capacity as a near-term bottleneck they cannot solve solely with giant campuses. Smaller, distributed deployments are a practical hedge against the long lead times of the megascale approach.